OpenAI has officially launched GPT-4.5, a new AI model known by the codename ‘Orion’. OpenAI unveils GPT-4.5 ‘Orion’ as its latest AI model, claiming improvements in accuracy and reasoning. This latest version is the company’s most advanced model to date, trained with significantly higher computing power and data than its predecessors.
Initially, OpenAI stated in its white paper that GPT-4.5 is not considered a “frontier AI model.” However, hours after its release, this line was removed. The updated document no longer includes this classification, raising speculation about OpenAI’s stance on GPT-4.5’s capabilities.
Availability and Pricing
Subscribers to ChatGPT Pro, OpenAI’s premium plan costing $200 per month, can access GPT-4.5 immediately. API developers with paid accounts also have access starting today. Users of ChatGPT Plus and ChatGPT Team are expected to receive the model next week, according to an OpenAI spokesperson.
OpenAI has priced GPT-4.5’s API at $75 per million input tokens and $150 per million output tokens. In comparison, GPT-4o, a widely used model, costs only $2.50 per million input tokens and $10 per million output tokens. Due to its high operating costs, OpenAI is assessing whether it will continue offering GPT-4.5 through its API in the long run.
Performance Gains and Limitations
GPT-4.5 was built using the same scaling approach as previous models. Larger datasets and more computing power have typically led to performance improvements in AI models. However, while GPT-4.5 shows some advancements, its results on AI benchmarks suggest diminishing returns from scaling.
On OpenAI’s SimpleQA benchmark, which evaluates factual accuracy, GPT-4.5 outperformed GPT-4o and some reasoning models like o1 and o3-mini. However, it fell behind DeepSeek’s latest models and OpenAI’s own advanced reasoning AI. Perplexity’s Deep Research model also scored higher in factual accuracy.
GPT-4.5 demonstrated mixed results on coding benchmarks. On the SWE-Bench Verified test, which assesses accuracy in coding tasks, GPT-4.5 matched GPT-4o but performed worse than Deep Research and Anthropic’s Claude 3.7 Sonnet. However, on OpenAI’s SWE-Lancer benchmark, which evaluates AI’s ability to develop full software features, GPT-4.5 outperformed GPT-4o but still lagged behind Deep Research.
Strengths in Human Interaction
OpenAI unveils GPT-4.5 ‘Orion’, stating that it has a deeper understanding of human intent and emotional intelligence. Despite its inconsistencies in technical benchmarks, GPT-4.5 showed improvements in conversational ability. OpenAI claims the model has “higher emotional intelligence” and provides more natural and socially appropriate responses. In a test where models were asked to console someone who failed a test, GPT-4.5 offered a more empathetic and supportive response compared to GPT-4o and o3-mini.
Moreover, GPT-4.5 performed well in creative tasks. When prompted to generate an SVG image of a unicorn, it was the only model that produced a recognizable image.
Future of AI Scaling
GPT-4.5 is seen as a crucial experiment in the evolving AI landscape. OpenAI has acknowledged that scaling up data and computing power alone may no longer yield exponential improvements. OpenAI unveils GPT-4.5 ‘Orion’, but experts question whether it truly surpasses previous models in overall performance. Industry experts have raised concerns that AI models are reaching the limits of conventional training methods.
In response, AI companies are shifting focus toward reasoning models, which take longer to process tasks but offer more reliable outputs. OpenAI plans to merge its GPT and “o” series models, beginning with GPT-5, expected later this year. GPT-4.5, despite its high costs and delayed launch, is likely a step toward more advanced AI systems in the near future.
